Research On Learning To Rank With Matrix Factorization And Deep Neural Network

Together with the boom of the Internet industry,people gradually transit from past age of information shortage to present age of information overload.Personalized recommendation system could present personalized contents to users by analyzing users'interests,exploring users' potential preferences,thus effectively solving the problem of information overload.Higher recommendation preciseness is proportional to better user experience.By adopting the machine learning technology to predict users' preference,learning to rank method could effectively promote the preciseness of the recommendation system.As a result of the fast development of deep learning technology in recent years,deep learning makes greater improvement on prediction precision in comparison with common machine learning method.Due to the complicated network structure,deep learning could fit more non-linear relation among characteristic and show stronger information expression ability.However,complicated network structure leads to the variety of model parameters,which could cause problems like gradient disappearance and inadequate training during the training process.Accordingly,this thesis raises a training framework which based on matrix decomposition,clustering and Wide&Deep product-based neutral network and expects to effectively solve problems in deep learning model caused by complicated network structure and improve model prediction precision.The thesis elaborately introduces the research status and basic principle of applying learning to rank to the recommendation system and proposes a new learning to rank framework targeted at shortcomings of existing methods.Different from common learning to rank methods,it firstly uses the matrix decomposition model to embedding users and products,then outputs the implicit vectors of users and products as the feature input of deep neural network.It secondly takes the clustering method to make a clustering classification for the implicit vectors of users and products,and outputs a coarse-grained classification for the behavioral data of users and products as the feature input of network.In terms of model network structure,the thesis introduces product layer network structure and puts forward WDPN(Wide&Deep Product-based Neural Network)model for training so that the model could more fully learn the feature input.Finally,model scores in the training can predict users' preference for products and generate the recommendation list for users.In order to verify the validity of learning to rank framework,the thesis compares the difference of effects between this framework and existing common frameworks according to two international public datasets.At the same time,the thesis makes a multi-dimensional analysis on the hyper-parameters in the framework.As proved by the test results,the learning to rank framework raised in this thesis effectively improves model prediction precision and recommendation system precision comparing with common learning to rank models in evaluation indicators(offline AUC,MAP@20).